Quantitative Techniques for Competition and Antitrust Analysis_11 ppt

338 6. Identification of Conduct econometric model, but it can also be helpful when collecting other evidence in a given case (e.g., documentary evidence). On the other hand, such an observation may concern us since we noted earlier that on occasion cartels have often resulted in relatively less variation in prices, perhaps because of stability concerns. As Corts (1999) noted, a different model of collusion would have different implications for observed collusive prices. 6.2.4.2 Identification of Pricing and Demand Equations in Differentiated Markets In a fashion entirely analogous to the homogeneous products case, the identification of conduct generally requires that the parameters of the demand and pricing equa- tions are identified. Even if demand rotation can also be used to identify conduct in differentiated industries in the same way as is done for homogeneous products, demand does need to be estimated to confirm or validate assumptions. This presents a challenge because a differentiated product industry has one demand curve and one pricing function for each of the products being sold. In contrast, in the homogeneous product case, there is only one market demand and one market supply curve that need to be estimated. Now we will need to estimate as many demand functions as there are products and also as many pricing equations as there are products. Iden- tification naturally becomes more difficult in this case and some restrictions will have to be imposed in order to make the analysis tractable. We discuss differentiated product demand estimation extensively in chapter 9. A general principle for identification of any linear system of equations is that the number of parameter restrictions on each equation should be equal to, or greater than, the number of endogenous variables included in the equation. A normalization restriction is always imposed in the specification of any equation so in practice the number of additional restrictions must equal or be more than the number of endogenous variables less one. 49 This is equivalent to saying that the restrictions must be equal to or more than the number of endogenous variables on the “right- hand side” of any given equation. The total number of endogenous variables is also the number of equations in the structural model. This general principle is known as the “order condition” and is a necessary condition for identification in systems of linear equations. It may, however, not be sufficient in some cases. Previously, we encountered the basic supply-and-demand two-equation system, where we had two structural equations with two endogenous variables: price and quantity. In that case we needed the normalization restrictions and then at least one parameter restriction for each equation for identification. We obtained the parameter restrictions from theory: variables that shifted supply but not demand were needed in the equations to identify the demand equation and vice versa (these exclusion restrictions are imposed by restricting values of the parameters to zero). A more technical discussion 49 The normalization restriction is usually imposed implicitly by not placing a parameter on whichever one of the endogenous variables is placed on the left-hand side of an equation. 6.2. Directly Identifying the Nature of Competition 339 Table 6.7. Nature of competition in the U.S. car market. Auto % in production Real auto quality-adjusted Sales Quantity Year (units) price/CPI prices revenues ($) index 1953 6.13 1.01 — 14.5 86.8 1954 5.51 0.99 — 13.9 84.9 1955 7.94 0.95 2.5 18.4 117.2 1956 5.80 0.97 6.3 15.7 97.9 1957 6.12 0.98 6.1 16.2 100.0 Source: Bresnahan (1987). of identification of demand and pricing equations in markets with differentiated products is provided in the annex to this chapter (section 6.4), which follows Davis (2006d). 6.2.4.3 Identification of Conduct: An Empirical Example When conduct is unknown, we will want to assess the extent to which firms take into account the consequences of pricing decisions on other products when they price one particular good. In this case, one strategy is to estimate the reduced form of the structural equations and retrieve the unknown structural parameters by using the correspondence between reduced-form and structural parameters derived from the general structural specification.Assuming that the demand parameters are identified and marginal costs are constant, we will need enough demand shifters excluded from a pricing equation to be able to identify the conduct parameters (see Nevo 1998). In particular, we will need as many exogenous demand shifters in the demand equation as there are products produced by the firm. Although identification of conduct is therefore technically possible, in practice it may well be difficult to come up with a sufficient number of exogenous demand and cost shifters. An early and important example of an attempt to identify empirically the nature of competition in a differentiated product market is provided by Bresnahan’s (1987) study of the U.S. car industry in the years 1953–57. Bresnahan considers the prices and number of cars sold in the United States during those years and attempts to explain why in 1955 prices dropped significantly and sales rose sharply. In particular, he tests whether this episode marks a temporary change of conduct by the firms from a coordinated industry to a competitive one. The data that Bresnahan (1987) is trying to explain are presented in table 6.7. The important feature of the data to notice is that it is apparent that 1955 was an atypical year with low prices and high quantities. Real prices fell by 5%, quantity increased by 38%, and revenues increased by 32%. To begin to build a model we must specify demand. Bresnahan specifies demand functions where each product’s demand depends on the two neighboring products in 340 6. Identification of Conduct terms of quality: the immediately lower-quality and the immediately higher-quality product. He motivates his demand equation using a particular underlying discrete choice model of demand but ultimately his demand function takes the form, q i D ı Ä P j  P i x j  x i  P i  P h x i  x h  ; where P and x stand for price and quality of the product and h, i, and j are indicators for products of increasing quality. Quality is one dimensional in the model, but captures effects such as horsepower, number of cylinders, and weight. Note that, all else equal, demand is linear in the prices of the goods h, i, and j and that given a price differential the cross-price slopes will increase with a decrease in the difference in quality, x. In this rather restrictive demand model there is only a single parameter to estimate, ı. To build the pricing equations, he assumes a cost function where marginal costs are constant in quantity produced but increasing in the quality of the products so that x j > x i > x h for products j , i , and h. These assumptions imply that the whole structure can be considered as a particular example of a model where demand is linear in price and marginal costs are constant in output. By writing a linear-in- parameters demand equation, where q i D ˛ i0 C˛ ii p i C˛ ij p j C˛ ih p h , we can see that for fixed values of the quality indices, x i , x j , and x h , the analysis of a pricing game using Bresnahan’s demand model can be incorporated into the theoretical structure we developed above for the linear demand model where the parameters in the equation are in fact functions of data and a single underlying parameter. (More precisely, we studied the linear demand model with two products above and we will study the general model in chapter 8.) Specifically, the linear demand parameters are of the form, ˛ ii Dı  1 x j  x i C 1 x i  x h à ; ˛ ij D ı  1 x j  x i à ; ˛ ih D ı  1 x i  x h à : Bresnahan estimates the system of equations by assuming first that there is Nash competition so that the matrix  describes the actual ownership structure of products (i.e., there is no collusion). Subsequently, he estimates the same model for a cartel by setting all the elements of the  matrix to 1 so that profits are maximized for the entire industry. He can then use a well-known model comparison test called the Cox test to test the relative explanatory power of the two specifications. 50 Bresnahan 50 Wehave shown thatthe two modelsBresnahan writesdownare nestedwithin a single family ofmodels so that we can follow standard testing approaches to distinguish between the models. In Bresnahan’s case he chooses to use the Cox test, but in general economic models can be tested between formally irrespective of whether the models are nested or nonnested (see, for example, Vuong 1989). 6.3. Conclusions 341 P MC P MC 1234 5 6 1234 5 6 (a) (b) Figure 6.5. Expected outcomes under (a) competition and (b) collusion. Source: Authors’ rendition of figure 2 in Bresnahan (1987). (a) Under competition, products with close substi- tutes produced by rivals get very low markups over MC. (b) Under collusion, close substitutes produced by rivals get much higher markups over MC. concludes that the cartel specification explains the years 1954 and 1956 while Nash competition model explains the data from 1955 best. From this, he concludes that 1955 amounted to a temporary breakdown of coordination in the industry. Intuitively, Bresnahan is testing the extent to which close substitutes are con- straining each other. If the firm maximizes profits of the two products jointly, there will be less competitive pressure than in the case where the firm wants to maximize profits on one of the products only and therefore ignores the negative consequences of lower prices on the sales of the close substitute product. Thus, in figure 6.5, if close substitute products 2 and 3 are owned by rivals, then they will have a low markup under competition but far higher markups under collusion. Given his assumptions about costs and the nature of demand, Bresnahan finds that the explanation for the drop in price during 1955 is the increase in the level of competition of close substitutes in the car market. The demand shifters that helped identify the parameter estimates are presented in table 6.8 as well as the accounting profits of the industry. The accounting profits, however, are not consistent with Bresnahan’s theory, as he notes. If firms are coor- dinating in the years 1954 and 1956, industry profits should be higher than in 1955 when they revert to competition. Bresnahan’s response is that accounting profits are not representative of economic profits and are not to be relied upon. We must there- fore make a decision in this case about whether to believe the accounting measures of profitability or the econometric analysis. In other cases, one might hope each type of evidence allows us to build toward a coherent single story. 6.3 Conclusions  Structural indicators such as market shares and concentration levels are still commonly used for a first assessment of industry conduct and performance, 342 6. Identification of Conduct Table 6.8. Demand and cost shifters of the car market in the United States 1953–57. Per capita disposable income Durable ‚ …„ ƒ Interest expenditures Accounting Year Level Growth rates (nonauto) profits ($) 1953 1,623 — 1.9 14.5 2.58 1954 1,609 0.9% 0.9 14.5 2.25 1955 1,659 3.0% 1.7 16.1 3.91 1956 1,717 3.5% 2.6 17.1 2.21 1957 1,732 0.9% 3.2 17.0 2.38 Source: Bresnahan (1987). although they are not usually determinative in a regime applying an effects- based analysis of a competition question. The fact that they are not determi- native does not mean market shares are irrelevant, however, for a competition assessment and many authors consider they should carry some evidential weight.  Developments in static economic theory and the availability of data have shown that causality between market concentration and industry profitability cannot be easily inferred. However, economic theories built on dynamic mod- els do frequently have a flavor of considerable commonality with the older SCP literature. For example, Sutton (1991, 1998) emphasizes that prices are indeed expected to be a function of market structure in two-stage games where entry decisions are made at the first stage and then active firms compete in some way (on prices or quantities) or collude at a second stage.  The broad lesson of game theory is that quite detailed elements of the com- petitive environment can matter for a substantial competition analysis. The general approach of undertaking a detailed market analysis aims at directly identifying the nature of competition on the ground and therefore the likely effects of any merger or alleged anticompetitive behavior.  Technically, the question of identification involves asking the question of whether two models of behavior can be told apart from one another on the basis of data. The hard question in identification is to establish exactly which data variation will be helpful in moving us to a position where we are able to tell apart some of our various models. The academic analysis of identification tends to take place within the context of econometric models, but the lessons of such exercises typically move directly across to inform the kinds of evidence that competition authorities should look for more generally such as evidence from company documents. 6.4. Annex: Identification of Conduct in Differentiated Markets 343  The degree to which firms are reactive to changes in demand conditions in the market can provide direct evidence of the extent of a firm’s market power. Formal econometric models can use the methods involving the estimation of conduct parameters in structural models to determine whether the reac- tions of firms to changes in prices are consistent with competitive, competing oligopoly, or collusive settings. However, the more general lesson is that changes in the demand elasticity can provide useful data variation to identify conduct. For example, we might (at least conceivably) find documentary evi- dence suggesting that firms’ pricing reactions accommodate prices in a fash- ion consistent with a firm’s internal estimates of market demand sensitivities (rather than firm demand sensitivities).  We examined identification results for both homogeneous product markets and also subsequently differentiated products markets. Analysis of identification in the former case suggests that demand rotators are the key to identifica- tion. In the differentiated product case, the results suggest that (i) examining the markups of close-substitute but competing products may be useful and (ii) examining the intensity with which demand and cost shocks to neighboring products are accommodated may sometimes be helpful when understanding the extent of coordination in a market.  In examining the likelihood of collusion, one must assess whether the neces- sary conditions for collusion exist. Following Stigler (1964), those are agree- ment, monitoring, and enforcement. The assessment of each of these con- ditions will typically involve a considerable amount of qualitative evidence although a considerable amount of quantitative evidence can be brought to bear to answer subquestions within each of the three conditions. For exam- ple, the European Commission examined the extent to which transaction prices were predictable given list prices to examine market transparency in the Sony–BMG case.  In addition to qualitative analysis of the factors which can affect the likelihood of collusion, it is sometimes possible and certainly desirable to develop an understanding of the incentives to compete, collude, and also to defect from collusive environments. 6.4 Annex: Identification of Conduct in Differentiated Markets In this annex we follow Davis (2006d), who provides a technical discussion of identification of (i) pricing and demand equations in differentiated product markets and (ii) firm conduct in such markets. In particular, we specify in more detail our example of a market with two firms and two differentiated products. Define the 344 6. Identification of Conduct marginal costs of production which depend on variables w such as input costs to be independent of output so that " c 1t c 2t # D "  0 1 0 0 0 2 #" w 1 t w 2 t # C " u 1t u 2t # : Similarly, suppose that demand shifters depend on some variables x such as income or population size which affect the level of demand for each of the products: " ˛ 01t ˛ 02t # D " ˇ 0 1 0 0ˇ 0 2 #" x 1 t x 2 t # C " " 1t " 2t # : Then linear demand functions for the two products can be written as " q 1 q 2 # D " ˛ 01 ˛ 02 # C " ˛ 11 ˛ 12 ˛ 21 ˛ 22 #" p 1 p 2 # ; while the pricing equations derived from the first-order conditions are " ˛ 11  12 ˛ 21  21 ˛ 12 ˛ 22 #" p 1  c 1 p 2  c 2 # C " q 1 q 2 # D 0: The full structural form of the system of equations is 2 6 6 6 4 ˛ 11  12 ˛ 21 10  21 ˛ 12 ˛ 22 01 ˛ 11 ˛ 12 10 ˛ 21 ˛ 22 01 3 7 7 7 5 2 6 6 6 4 p 1 p 2 q 1 q 2 3 7 7 7 5  2 6 6 6 4 ˛ 11  0 1  12 ˛ 21  0 2 00  21 ˛ 12  0 1 ˛ 22  0 2 00 00ˇ 0 1 0 000ˇ 0 2 3 7 7 7 5 2 6 6 6 4 w 1 t w 2 t x 1 t x 2 t 3 7 7 7 5 D 2 6 6 6 4 v 1t v 2t v 3t v 4t 3 7 7 7 5 or, more compactly in matrix form, Ay t C Cx t D v t ; where the vector of error terms is in fact a combination of the cost and demand shocks of the different products, 2 6 6 6 4 v 1t v 2t v 3t v 4t 3 7 7 7 5 D 2 6 6 6 4 ˛ 11  12 ˛ 21 00  21 ˛ 12 ˛ 22 00 0010 0001 3 7 7 7 5 2 6 6 6 4 u 1t u 2t " 1t " 2t 3 7 7 7 5 : Following our usual approach, this structural model can also be written as a reduced- form model: y t DA 1 Cx t C v t D ˘x t C v t : 6.4. Annex: Identification of Conduct in Differentiated Markets 345 The normalization restrictions are reflected in the fact that every equation has a 1 for one of the endogenous variables. This sets the scale of the parameters in the reduced form so that the solution is unique. If we did not have any normalization restrictions, the parameter matrix ˘ could be equal to A 1 C or equivalently (in terms of observables) equal to .2A/ 1 2C . In our structural system we have four equations and four endogenous variables. Our necessary condition for identification is therefore that we have at least three parameter restrictions per equation besides the normalization restriction. In gen- eral, in a system of demand and pricing equations with J products, we have 2J endogenous variables. This means that we will need least 2J 1 restrictions in each equation besides the normalization restriction imposed by design. There are exclusion restrictions that are imposed on the parameters that come from the specification of the model. First, we have exclusions in the matrix A which are derived from the first-order conditions. Any row of matrix A will have 2J elements, where J is the total number of goods. There will be an element for each price and one for each quantity of all goods. But each pricing equation will have at most one quantity variable in, so that for every equation we get J  1 exclusion restrictions immediately from setting the coefficients on other good’s quantities to 0. Second, the ownership structure will provide exclusion restrictions for many models. Specifically, in the pricing equations, there will only be J i parameters in the row, where J i D P J j D1  ij is the total number of products owned by firm i (or, under the collusive model, the total number of products taken into account in firm i’s profit-maximization decision). The implication is that we will have J  J i restrictions. Third, in each of the demand equations in matrix A, we also have J 1 exclusion restrictions as only one quantity enters each demand equation (together with all J prices); the parameters for the other J  1 quantities can be set to 0. Fourth, we have exclusion restrictions in matrix C which come from the existence of demand and cost shifters. Demand shifters only affect prices through a change in the quantities demanded and do not independently affect the pricing equation. Similarly, cost shifters play no direct role in determining a consumer’s demand for a product; they would only affect quantity demanded through their effect on prices. Those cost and demand restrictions are represented by the zeros in the C matrix. Define k D as the total number of demand shifters and k C as the total number of cost shifters. For each of the pricing equations in C we have k D exclusion restrictions because none of the demand shifters affect the pricing equation directly. Similarly, for each of the demand equations we have k C exclusion restrictions since none of the cost shifters enter the demand equations. Additionally, even though any row in matrix C will have as many elements as there are exogenous cost variables and demand shifters, there will only be as many new parameters in a pricing equation as there are cost shifters in that product’s pricing 346 6. Identification of Conduct equation. Similarly, there will only be as many new parameters in the demand equation as there are demand shifters in that product’s demand equation. In addition to the exclusion restrictions we have just described, there are also cross-equation restrictions that could be imposed on the model. Cross-equation restrictions arise, for example, when we have several products produced by a firm. In that case, since prices are set to maximize joint profits for the firm, their pricing equations will be interdependent for that reason. Theory predicts that the way the demand of product j affects product i’s pricing equation is not independent of the way the demand of product i affects product j ’s pricing equation. This gives rise to potential cross-equation restrictions. For example, the matrix A we wrote down has a total of sixteen elements but in fact it has only four structural parameters. We could impose that the reduced-form parameters satisfy some of the underlying structural (theoretical) relations. For instance, the first elements of rows 1 and 3 are the same parameter with opposite signs. This could be imposed when determining whether the structural parameters are in fact identified from estimates of the reduced-form parameters. The more concentrated the ownership of the products in the market the more cross-equation restrictions we will have, but the fewer exclusion restrictions we will have since we will have fewer zero elements of . In addition, we will need more exclusion restrictions in each pricing equation to identify all the demand parameters that will be included. 7 Damage Estimation The estimation of damages has been one field within antitrust economics where quantitative analysis has been used profusely. Most of the work has been done in countries where courts set fines or award compensation payments that are based on the estimated damages caused by infringing firms. Effective deterrence using fines, as distinct from, say, criminal conviction of individuals, requires that imposed fines be at least as high as the expected additional profits of firms that would emanate from the behavior to be deterred. Expected profits can be difficult to measure and in cartel cases they are currently often approximated by the damages caused to affected customers. This chapter describes the issues investigators confront in estimating the damages caused by the exercise of market power by cartels. We also briefly discuss damage calculations from abuses by a single firm. 7.1 Quantifying Damages of a Cartel A presumption of antitrust law is that cartels are bad for consumers. Both antitrust agencies and customers see that cartels increase prices and reduce the supply avail- able on the market. For this reason, cartels are illegal in most jurisdictions. For example, the Sherman Act in the United States, Article 81 in the EU and Chapter 1 of the Competition Act (1998) in the United Kingdom each prohibit firms from coordinating in order to reduce competition. Nonetheless, because cartels that work can be very profitable there is a temptation to collude when the conditions in the market make it possible. Illegality per se is not enough of a deterrent when it is not accompanied by at least the potential for a punishment that will hopefully wipe out the expected benefits of participating in a cartel. Cartels are increasingly pun- ished with substantial fines and in some jurisdictions including the United States and the United Kingdom some cartel behavior is a criminal offense. 1 For a fine to 1 Section 188 of the U.K. EnterpriseAct 2002 introduced a criminal offense for collusion in the United Kingdom. It says, for example, that an individual is guilty of an offense if he “dishonestly agrees with one or more other persons” to, in particular, directly or indirectly fix prices. Note that the word “dishonestly” qualifies the word “agrees” so that not all agreements to fix prices are immediately dishonest and hence not all cartel offenses are criminal offenses. The term dishonest is frequently used under other parts of criminal law and so has clear legal status relating both to whether a person’s actions were honest [...]... intuitive and may even be fairly accurate In other contexts, there are cases where a purely statistical approach to forecasting can sometimes perform better than building an economic model and basing the forecast on that Either approach requires assumptions For example, the raw form of the before -and- after methodology implicitly assumes that market conditions are unchanged since if demand and supply... cartel was 7.1.2.2 Before and After The “before -and- after” methodology uses the historical time series of the prices of the cartelized goods as the main source of information It looks at the prices before and after the cartel and compares them with the prices that prevailed during the cartel The damages are then calculated as the difference between the cartel prices and the prices under competition multiplied... environment Information on costs and demand parameters allow us to undertake simulations under other simple competitive frameworks Consider, for example, an industry with two differentiated products produced by firms competing in prices and facing linear differentiated product demands and constant (in quantity) marginal costs of production The structural form of the “supply” (i.e., pricing) and demand equations... function and on the price elasticities of both demand and supply In particular, for any given increase in costs and shift in the supply curve, the pass-on will be larger when the supply curve is more elastic and when the demand curve is more inelastic Figure 7.8 compares the pass-on rate for a competitive market with an elastic demand to that with a competitive market with an inelastic demand We do... demand curve at the point where supply and demand intersect and in each case seeing what happens when supply shifts The graph shows that when demand is inelastic, the quantity effect is smaller and so the increase in marginal cost will be passed on to consumers to a larger extent than when demand is elastic, all else equal An elastic demand makes price increases less profitable for the producer and. .. what would have happened during the cartel period if competition had prevailed 7.1.2.3 Multivariate Approach One can attempt to overcome the criticisms of the simplest version of the “beforeand-after” method by taking into account changes in demand and supply conditions By running a reduced-form regression of the price level on demand and cost factors 7 For a good discussion of the overcharge estimation... uncommon to perform more elaborate regressions, it is important that the results of the reduced form be at least compared with alternative specifications to check for robustness A second multivariate approach is to forecast the “but for price that would have prevailed during the cartel period absent the conspiracy Using pre-cartel and postcartel data the effect of the determinants of demand and cost shifters... firm to raise prices to its customers to compensate for the effect of the higher costs on profits As in the case for the direct damages, there are two basic approaches to estimating the pass-on effect: a reduced-form approach and a structural model approach and we discuss each in turn 7.1.3.2 Reduced-Form Approach for Calculating the Pass-On The reduced-form approach measures the effect of an increase... The reduced form does not control for endogenous variables such as quantity since those would only enter a structural form if prices and quantities are determined in equilibrium 7.1 Quantifying Damages of a Cartel 371 7.1.3.3 Structural Approach for Calculating the Pass-On The structural approach specifies a model of competition in the downstream market It must specify a demand function and a pricing... Those values of the parameters can be used to predict the “but for price during the cartel The difference between the actual price and the predicted price provides a prediction of the overcharge As opposed to the simple before-andafter method, forecasting the price by running multivariate regression can allow for changes in the demand and supply conditions However, it assumes that the structural relation . of the “before- and- after” method by taking into account changes in demand and supply conditions. By running a reduced-form regression of the price level on demand and cost factors 7 For a good. was. 7.1.2.2 Before and After The “before -and- after” methodology uses the historical time series of the prices of the cartelized goods as the main source of information. It looks at the prices before and. matter for a substantial competition analysis. The general approach of undertaking a detailed market analysis aims at directly identifying the nature of competition on the ground and therefore